Enhancing accuracy of prediction of critical heat flux in Circular channels by ensemble of deep sparse autoencoders and deep neural Networks

Accurate prediction of Critical Heat Flux (CHF) is essential for ensuring safety and economic efficiency of water-cooled reactors and two-phase flow boiling heat transfer systems. However, the lack of a deterministic theory for CHF prediction remains a significant challenge in the thermal engineering domain. This has led to the development of numerous prediction models based on various CHF experimental data, with no single universally acceptable model covering the wide range of flow conditions encountered in practice. In this paper, we explore the use of a comprehensive CHF experimental dataset in conjunction with artificial intelligence techniques to predict CHF in vertical tubes, contributing to the ongoing efforts to address this critical issue. The proposed method stands on the collection of comprehensive CHF experimental data from various sources, covering a wide range of operating conditions (pressure of 100 – 21,197 kPa, hydraulic diameters of 1 – 44.7 mm, mass fluxes of 10 – 20,910 kg/m²s, inlet-subcooling of 0.6 – 3,555 kJ/kg, heated lengths of 9 – 6,000 mm and critical qualities of −0.494 – 0.981), and is based on a new prediction model for the prediction of the CHF. Specifically, the prediction model consists of an ensemble of deep sparse autoencoders (AEs) used as a base-learner to extract robust features from the input data and a deep neural network (DNN) built on top of the ensemble of deep sparse AEs for use as a meta-learner to predict the CHF. The proposed method is validated on the collected CHF data and the obtained results show a substantial improvement in CHF prediction accuracy, outperforming standalone and other state of-the-art machine learning models. This innovative approach offers a notable improvement in CHF prediction, potentially contributing to the development of more reliable and efficient nuclear reactors.